Lower-than-expected throughput using Triton with GPU and CPU

[Manutea]

Hi,

I am using Triton Server Version 2.39.0 with a GPU and a CPU each with an instance of a DNN model.
I notice that the throughput is lower than if I were using a single GPU.
Perhaps I have configured my model incorrectly or the options I have added with the perf-analyser command are not good.

Here is my config.pbtxt

name: "googlenet"
platform: "onnxruntime_onnx"
max_batch_size: 256
input [
  {
    name: "input"
    data_type: TYPE_FP32
    dims: [3, 224, 224]
  }
]
output [
  {
    name: "output"
    data_type: TYPE_FP32
    dims: [1000]
  }
]
instance_group [
  {
    count: 1
    kind: KIND_GPU
  },
  {
    count: 1
    kind: KIND_CPU
  }
]

The command line :

perf_analyzer --async -b $batch -m googlenet --concurrency-range $concurrency

And the plot (the numbers on the dots are the batch size) :